Propagation depth-aware knowledge graph neural networks for recommendation

In recent years, recommender systems based on graph neural networks (GNNs) and knowledge graphs (KGs) have been designed to consider various aspects of users and items in order to provide accurate recommendations. Numerous studies have demonstrated that leveraging the rich semantics of KGs effectively enhances recommendation performance, alleviates data sparsity issues, and provides interpretable recommendations. Many contemporary approaches represent users and items as undifferentiated entities, implying that their embeddings spread according to a single, uniform propagation scheme. In practice, however, people exhibit a variety of characteristic facets, and the degree to which each facet influences others differs from one individual to another. Consequently, carefully evaluating and weighting these attributes is essential for enhancing recommendation accuracy. Thus, we introduce a Propagation Depth-aware Knowledge Graph Neural Networks (PDKGNN) for recommendation systems. First, PDKGNN draws on both the knowledge graph and the user-item interaction data to harvest auxiliary information, then merges this information into user and item representations via a Convolutional Neural Networks (CNN). Second, we split each user’s and item’s attributes into separate embedding channels, enabling us to capture a detailed propagation process in which every attribute embedding travels through its own depth. Third, we implement an attention mechanism aggregation scheme that highlights particular attribute facets and blends the diverse embeddings according to learned attention weights. Empirical evaluations on three real-world datasets (Yelp, MovieLens-1M, and Book-Crossing) demonstrate that PDKGNN exceeds a variety of leading baselines, yielding an average performance boost of up to 7.23%.